MEDVSE is a collaborative research project that develops efficient deep learning methods for estimating vital signs (heart rate and blood oxygen saturation) from smartphone videos of fingertips. This work was published as “Efficient Deep Learning-based Estimation of the Vital Signs on Smartphones”.
Key Innovations
- Smartphone-based Vital Signs: Non-invasive estimation of heart rate and SpO2 using only a smartphone camera
- Custom Dataset: Created MTHS (Medical Taha-Mahdi Heart rate and SpO2) dataset for training and evaluation
- Efficient Architecture: Optimized neural network design for mobile deployment
- Real-time Processing: Capable of processing video streams for immediate vital sign estimation
- High Accuracy: Achieved state-of-the-art results for smartphone-based vital sign monitoring
Technical Approach
The project uses a deep learning approach to process RGB signals extracted from fingertip videos:
- RGB signal extraction from video frames
- Signal preprocessing and feature extraction
- Deep neural network for vital sign estimation
- Validation against medical-grade reference devices
Dataset
As part of this project, we created and released the MTHS dataset:
- Fingertip videos from multiple subjects
- Ground truth data from medical devices
- RGB signals sampled at 30Hz
- Heart rate and SpO2 measurements sampled at 1Hz
Applications in Healthcare
- Remote Patient Monitoring: Allow patients to monitor vital signs at home
- Telehealth: Support remote clinical assessment
- Personal Health Tracking: Enable individuals to track their vital signs regularly
- Resource-Limited Settings: Provide vital sign monitoring where medical equipment is limited
- COVID-19 Monitoring: Support home monitoring of respiratory parameters
This collaborative project with Mahdi Farvardin demonstrates how deep learning can transform smartphones into powerful tools for health monitoring, making vital sign measurement more accessible worldwide.